Visual Jev is here Following the community's approach

Replicated Jev's fixed-label sampling approach with a 4-bit multimodal Qwen model running locally on a 16GB M4 MacBook.

The original post, translated

Visual Jev is here

Following the community's approach, mapping candidate results to fixed labels, sampling directly after prefill, limiting output tokens, using a multimodal qwen-3.5-0.8B 4bit quantized deployment locally, running smoothly on a 16g m4 mbp

Disclaimer: this is just a small reproduction of jev's form at the infer layer, definitely not as efficient as various mature inference frameworks, and definitely not jev's actual principle, but as a small introduction to inference it's still pretty good

Demo as follows, achieving "Three-Color Goods Sorting" in 200ms

The post above is a machine translation from zh; the untranslated text is in the fold-out.

What the author linked

• github.com/hr98w/jev-visu…
• GitHub - hr98w/jev-visual: An educational Jev-like visual inference experiment on Apple Silicon: shared context, direct candidate scoring, and local visual demos. — An educational Jev-like visual inference experiment on Apple Silicon: shared context, direct candidate scoring, and local visual demos.

Engagement when collected

Numbers are a snapshot taken from X when the case was added to the library (schema v1, collected 2026-09-19); they will not match today.